Using a Cognitive Network Model of Moral and Social Beliefs to Explain Belief Change

Page created by Nelson Richards
 
CONTINUE READING
Using a Cognitive Network Model of Moral and Social
                                                     Beliefs to Explain Belief Change
                                                                                1,2
                                                             Jonas Dalege             and Tamara van der Does1,2
                                                       1
                                                        Equal authorship, order determined by universe splitter.
                                                  2
                                                      Santa Fe Institute, 1399 Hyde Park Rd, Santa Fe, NM 87501.
arXiv:2102.10751v2 [cs.SI] 2 Aug 2021

                                                                                       Abstract
                                                 Scepticism towards childhood vaccines and genetically modified food has grown
                                             despite scientific evidence of their safety. Beliefs about scientific issues are difficult to
                                             change because they are entrenched within many related moral concerns and beliefs
                                             about what others think. We propose a cognitive network model which estimates
                                             the relationships, dissonance, and randomness between all related beliefs to derive
                                             predictions of the circumstances under which beliefs change. Using a probabilis-
                                             tic nationally representative longitudinal study, we found support for our model’s
                                             predictions: Randomness of the belief networks decreased over time, for many partic-
                                             ipants their estimated dissonance related positively to their self-reported dissonance,
                                             and individuals who had high estimated dissonance of their belief network were more
                                             likely to change their beliefs to reduce this dissonance. This study is the first to
                                             combine a unifying predictive model with an experimental intervention and sheds
                                             light on dynamics of dissonance reduction leading to belief change.

                                        Introduction
                                        The World Health Organization (WHO) lists vaccination hesitancy as one of the ten
                                        greatest threats to global health [1]. Erroneous beliefs regarding vaccines, which are
                                        somewhat common in the US [2], can accelerate or even re-ignite the spread of diseases
                                        globally. In another recent report, the WHO also points out that around 45% of deaths
                                        among children under 5 years of age are linked to undernutrition [3]. Even though the
                                        scientific community has confirmed that currently approved Genetically Modified (GM)
                                        crops are safe and could provide higher yields [4], many US Americans are sceptical about
                                        this technology [5, 6]. Other beliefs inconsistent with the scientific consensus, such as
                                        climate change denial, can have similar detrimental consequences for society. We need
                                        to understand how these sceptical beliefs about scientific issues can be changed in order
                                        to develop successful public science communication.
                                            In this paper, we use a cognitive network model inspired by statistical physics to
                                        understand change in beliefs about GM food and childhood vaccines. We consider at-
                                        titudes towards GM food and childhood vaccines as networks of connected beliefs [7–9]

                                                                                          1
and use this model to precisely estimate the relationships, dissonance, and randomness
between all beliefs. Using data data from a longitudinal nationally representative study
with an educational intervention, we test if predictions derived from our cognitive net-
work model can explain under which circumstances individuals are more likely to change
their beliefs over time and shed light on the dynamic nature of dissonance reduction
leading to belief change. By combining a unifying predictive model with an experimental
longitudinal dataset, we expand upon the strengths of earlier investigations into science
communication and belief change dynamics, as we describe in the next paragraphs.
    Previous applied research on beliefs about GM food and childhood vaccines has found
that scepticism about their safety is shaped both by related moral beliefs (e.g., care for
others, concerns about the environment, importance of naturalness, and purity) and
perceived beliefs of trusted social groups, such as doctors or family members [10–16].
Studies focusing on changing these beliefs have therefore tried to vary the framing of
the factual information and the source of information, with mixed success [17–20]. This
literature sheds light on important related beliefs (both moral and social) as determinants
of beliefs about GM food and childhood vaccines. However, these empirical studies tend
to focus on specific interventions and populations [21] and do not draw on a unifying
model to understand the processes underlying belief change more generally.
    Mirroring findings from applied research on GM food and childhood vaccines, general
models of belief change have identified two important sets of factors as consistently
important for belief change (for a review, see [22]). First, individuals hold many related
personal beliefs, such as moral beliefs. In social psychology, the concept of dissonance
was developed to understand when and why individuals might change their beliefs when
they are incoherent with each other [23, 24]. Within this approach, incoherent beliefs
translate into feelings of dissonance and belief change if attention is paid to these beliefs.
Building on this concept of dissonance, more recent research has modelled the relationship
between personal beliefs using cognitive network models to predict belief dynamics [7–
9, 25, 26]. Second, individuals’ beliefs are shaped by their social networks. In statistical
physics, models of opinion dynamics can predict change over time within a social network
[27, 28]. These models, however, generally do not take into account that the beliefs held
by one’s social network do not directly influence one’s own beliefs. Instead, their influence
is mediated by how one perceives beliefs in one’s social network [29–32], which implies
that perceptions of beliefs in one’s social network provide information above and beyond
actual beliefs in one’s social network [33]. For example, a person might overestimate
how liberal their friends are, and thus become more liberal themselves, just because their
liberal friends voice their political position more firmly than their moderate friends.
    In recent years, a few belief change models were developed to focus specifically on
the interaction between related personal beliefs (e.g., moral beliefs) and beliefs about
one’s social network (social beliefs). In general, these models have either focused on
dissonance between all moral and social beliefs using the statistical physics concept of
energy [34], or on network imbalance between personal and social beliefs (e.g., belief A
is positively connected to beliefs B and C, but belief B and C are negatively connected)
[35–37]. Models based on dissonance were able to predict belief change using estimated

                                              2
energies from reported moral and social beliefs [38]. However, these models did not take
the network structure of related moral beliefs into account. Models based on network
imbalance were able to predict distributions of beliefs [36] and provide an explanation of
how minorities can convince majorities [35] but empirical tests of whether these model
can predict belief change are still lacking. In the following sections, we present our
cognitive network model before moving to the results section.
    Our cognitive network model integrates both moral and social beliefs to empirically
predict belief change dynamics. We extend the recent Attitudinal Entropy (AE) frame-
work [39], a model inspired by statistical physics which conceptualises individual overall
attitudes as networks comprising of different beliefs or “spins” (i.e., elements of the net-
work that can take different values) that are of binary nature. Within this framework,
individuals change their beliefs towards more consistency, so as to reduce their attitudinal
entropy (a measure of unpredictability of the whole belief network). We build upon this
model to include both moral and social beliefs and by generalising to beliefs that can take
any value between -1 (complete disagreement) and 1 (complete agreement). We discuss
the implications of having continuous beliefs in the results section. In the next three
paragraphs, we explain how each major concept in our model relates to psychological
constructs and belief change before discussing our empirical predictions.
    The main concepts of our cognitive network model, their proposed psychological
meaning, and the way we estimate them are listed in Table 1. Couplings between both
moral and social beliefs represent the strength and sign (positive or negative) of the
relationships between beliefs. These couplings are estimated using partial correlations
between nodes in the network. The strength of the couplings within the network de-
termines the probability of beliefs being in consistent states. A network with strong
couplings is more likely to have consistent beliefs, while a network with weak couplings
is more likely to have inconsistent beliefs. The sign of the couplings in the network
determines which belief states (spins) can be regarded as consistent. Let’s assume, for
example, that the belief that vaccines are safe and the belief that they are effective are
positively connected, and that the belief that vaccines are safe is negatively connected to
the belief that pharmaceutical companies are only interested in making money regardless
of patients’ health. The belief network would be highly consistent if the individual agrees
with the former two beliefs and disagrees with the later belief (or, conversely, disagrees
with the former two beliefs and agrees with the later belief).
    Energy can be understood as a formalisation of the psychological concept of disso-
nance. Dissonance refers to the actual inconsistency between beliefs and is translated
into felt dissonance if enough attention is given to these beliefs. According to classic psy-
chological theories and the AE framework, felt dissonance leads to belief change because
individuals want their beliefs to be in a consistent state [24, 39]. Energy is measured as
the sum of the products of each pair of beliefs (spins) and their relationship (coupling)
multiplied by -1. A consistent network has low energy and an inconsistent network has
high energy.
    Temperature can be interpreted using several psychological processes that increase
randomness and disorder, such as lack of attention and no thought directed to beliefs,

                                             3
Table 1: Overview of cognitive network model parameters within the statistical physics
framework, and their corresponding psychological constructs and methods of estimation.

 Statistical          Psychological construct             Estimation
 physics term
 Coupling ωij         Relationship between two be-        Partial correlation between bi and bj
                      liefs.                              controlled for all other beliefs.
 Energy H             Dissonance.                         Measures inconsistency (high energy) or
                                                          consistency (low energy) of beliefs given
                                                          estimated network structure. Sum of
                                                          weighted products of self-reported belief
                                                          scores, −ωij bi bj .
 Temperature 1/β      Subsumes several processes          Measures inverse of interdependence be-
                      that increase randomness and        tween beliefs. Average of the diagonal of
                      disorder of belief networks         the inverse covariance matrix of beliefs.
                      such as lack of attention and       In other words, average of belief-specific
                      no thought directed to the          scaling values of bi and bj which are es-
                      belief network.                     timated in order to transform ωij into
                                                          measured correlations.

with lower temperature corresponding to higher attention and thought. Temperature is
estimated using the average of scaling values transforming individual couplings (estimated
partial correlations) into measured correlations. Therefore, temperature relates to the
interdependence between beliefs: High measured correlations between beliefs result in
low estimated temperature, while low measured correlations between beliefs result in
high estimated temperature. Temperature influences belief change through its scaling of
the couplings. Lower temperatures increase the impact of couplings on the belief states,
while higher temperatures reduce the impact of the couplings.
    Relationships between different concepts in our cognitive network model can be ex-
pressed by Equations 1 and 2:
                                            X
                                    Hi = −       ωij bi bj                             (1)
                                             j:j6=i

    where Hi is the energy (dissonance) of a belief, bi is the value of belief i in the belief
network (moral or social), and ωij is the coupling (relationship between beliefs) between
bi and another belief in the network, bj . The energy of one belief Hi is the sum of this
beliefs’ individually weighted product with all other beliefs in the network. In this paper
we focus on the couplings, thus omitting the external field from the energy equation.
We discuss this and other assumptions in detail in the result section. The conditional
probability that a given belief will change its state from its current state to a different
state is
                                P (bi → b0i ) = 1/(1 + eβ∆Hi )                            (2)

                                              4
P ∆Hi =
where             Hi0 − Hi is the change in energy between the two belief states (Hi0 =
               0
− j:j6=i ωij bi bj ) and β is the inverse temperature of the network. The probability of
belief changing state from bi to b0i increases with (a) the difference in energy between the
new state and the current state (∆Hi ) and (b) the reduction in temperature of the whole
network (β). The lower the temperature, the higher the probability of belief change
towards a new state with lower energy.             P
    The whole belief network has energy H = i Hi , summing the energies of all the
beliefs in the network. We expect beliefs to change towards more consistency of the whole
network (Figure 1). This reflects the expected high probability of consistent networks
(low energy states) in equilibrium. As shown in Figure 1, there are multiple ways for
beliefs to change in order to achieve consistency. To note, our cognitive network model
proposes that beliefs are more likely to move to a consistent state if the network estimated
temperature is less than infinity (i.e., when at least some attention is directed at the
beliefs). As temperature increases, beliefs are less likely to move to a consistent state,
until there is no difference in probability between consistent and inconsistent states when
temperature is infinitely high.
    Three empirical predictions can be drawn from our cognitive network model. First,
in our cognitive network model, temperature measures the inverse of interdependence
between beliefs (Table 1) and therefore should relate to processes that decrease disorder
and randomness, such as attention to one’s beliefs. This leads to the prediction that
attention to beliefs should lead to lower estimated network temperatures. Second, our
model holds that belief network energies are a formalisation of dissonance. Individuals
with high energy should thus also have high feelings of dissonance. We therefore predict
that estimated energies should be positively related to self-reported feelings of dissonance.
Finally and most crucially, belief change is predicted to be more likely when individuals
try to achieve higher consistency between their beliefs in order to reduce these feelings of
dissonance. Our third prediction is therefore that individual energies should predict belief
change and belief change should be associated with a process of lowering energies. In
order to test these three predictions, we need longitudinal data on beliefs over time from
which we can estimate an empirical model of belief networks drawn from our cognitive
network model. Investigating the dynamics of dissonance reduction goes above and
beyond the usual investigations into cognitive dissonance, which typically only investigate
the consequences of inducing dissonance [40–42]. Here, we investigate how dissonance
interacts with receiving new information on a topic and whether such new information
leads to reconfiguration of one’s beliefs leading to lower dissonance.

Results
We used a nationally representative longitudinal study of beliefs about GM food and
childhood vaccines to test implications of our cognitive network model (see Methods for
details on the study design and questionnaire). This study included questions about both
moral beliefs related to the safety of each technology (e.g., GM food [Childhood vaccines]
are beneficial to children, GM food [Childhood vaccines] are part of our tradition) and

                                             5
(b)
                                             Consistent network

                                                                                     Higher probability
                              (a)
                     Inconsistent network
 Negative belief
 Negative coupling
                                                     (c)
 Positive coupling                           Consistent network
 Positive belief

                                                                      Lower energy

Figure 1: A belief network will change over time to achieve higher consistency. Network
(a) is inconsistent because it has similar beliefs connected by negative couplings and
opposite beliefs connected by positive couplings. The inconsistent network (a) will change
to achieve higher consistency, either towards (b) or (c). These changes reflect the higher
probability in equilibrium of a low energy network compared to a high energy network,
given a temperature less than infinity.

                                            6
social beliefs about their safety (e.g., % of medical doctors believe GM food [Childhood
vaccines] is [are] safe, % of my family and close friends believe GM food/Childhood
vaccines is [are] safe). We assessed these beliefs four times across three waves of data
collection (over an average of 30 days): once in the first and third wave and twice in
the second wave (before and after the intervention). In the second wave, we presented
individuals with an educational intervention about the safety of GM food and vaccines,
quoting reports from the National Academies of Sciences. A total of 979 individuals
participated in all three waves and answered all relevant questions for the study.

Network structure
To fit our cognitive network model to empirical data, we focused on changes in all moral
and social beliefs after removing variations explained by individual-level and time-level
differences (see Methods for details). We also made four key assumptions. First, we
assumed that Gaussian distributions are appropriate for our data (see empirical support
in Supplementary Materials Figure 1), which allowed us to estimate a Gaussian Graphical
Model (GGM). Second, our data represents equilibrium distributions. While individual
beliefs can change, we assumed a fixed distribution of all the beliefs in the belief network.
Third, in contrast to what is typically assumed for GGM models, we assumed that
individuals are not motivated to move to the global mean, because external fields vary
between individuals. We think that this assumption is appropriate, because individuals
have different dispositions for their beliefs, which makes it likely that they are motivated
to move to different belief states. Fourth, we assumed that couplings in the estimated
group-level belief networks are representative for the couplings at the individual level
(see empirical support in Supplementary Materials Figure 2)
    Our model can be specified in several different ways when fitted to empirical data.
We did not make any assumptions on (a) whether the couplings, the external fields,
and/or temperature vary over time and (b) whether the networks are sparsely or densely
connected. To investigate if our constructs vary over time and whether temperature drops
during the time course of the study, we fitted different specifications of our model on the
four time points. These specifications focused on constraining partial correlations to be
equal across time points, external fields (mean values) to be equal across time points,
and temperature to be equal across time points (see Methods for details on network
estimation). Additionally, we tested whether the data can be captured best by a dense
network (all beliefs are directly connected to all other beliefs) or a sparse network (some
beliefs are not directly connected). The results indicated that a sparse network with equal
partial correlations and equal external fields between time points and varying temperature
across time points fitted the data best. This implies that the network structure remained
constant throughout time, but that the interdependence between beliefs varied.
    The estimated group-level networks for beliefs regarding GM food and childhood vac-
cines are shown in Figure 2a and 2b. In both networks the moral and social beliefs were
connected to each other but formed two distinct clusters. Most beliefs were positively
connected but there were some negative connections as well. For example, there is a neg-
ative estimated coupling between the belief that scientists think GM food is safe and the

                                             7
GM food                                                    Childhood vaccines

      (a)                                 Gov
                                                                                    (b)                     OnE
                   GeP
                                                                                                    FaF           OnC
                                                                                                                               Agc           Inf
                        Jou
                                                 Sci                  FrC
                                                                                                    Med
                                    Med
            OnC                                                                           Sci
                                                                            All                                   Jou
                                                                                                                                                    Nat
                              FaF                                                                                                Chi
                  OnE                                   Agc
                                                                                                      Gov
                                                                            Com               GeP                                             Fam
                                                                Inf                                                            Env

                                                                                                                         All
                                           Nat                                                                                                        God
                        Tra                               Env                                                                          Cou
                                                                  Cou
                                                  Chi
                                                                                                                        Com
                                    God
                                                                                                          FrC
                                                          Fam                                                                          Tra

                                                                              Moral beliefs
                    Agc: Appropriate agencies approve                                   FrC: Free to choose
                    Com: Companies and individuals benefit                              God: God approves
                    Chi: Beneficial to children                                         All: All individuals benefit
                    Cou: Positive for country                                           Inf: Information is shared
                    Env: Beneficial to environment                                      Nat: Natural
                    Fam: Positive for family                                            Tra: Part of tradition

                                                                             Social beliefs
                                     OnE: Online experts                                Med: Medical doctors
                                     Fam: Family and friends                            OnC: Online community
                                     Gov: Governmenal agencies                          GeP: General public
                                     Jou: Journalists at favorite media                 Sci: US scientists

Figure 2: Belief networks for GM food (a) and childhood vaccines (b), N=979. Beliefs
include moral (orange nodes) and social beliefs (green nodes). Edges represent couplings
(partial correlations) between two beliefs controlled for all other beliefs. Blue (red) edges
represent positive (negative) couplings and the widths of the edges correspond to the
strength of the couplings. The strength of the couplings ranged from 0.02 (between the
beliefs "Chi" and "Fam") to 0.30 (between the beliefs "Med" and "Sci") for GM food and
from 0.02 (between the beliefs "Com" and "Jou") to 0.28 (between the beliefs "OnE"
and "OnC").

                                                                                   8
one that God approves of GM food. The GM food network was more densely connected
than the childhood vaccines network, indicating that beliefs toward childhood vaccines
were more independent from each other. This might indicate that individuals formed
more nuanced beliefs toward childhood vaccines than toward GM food. In addition to
estimating couplings between beliefs, we also estimated the overall temperature of the
networks over time.

Temperature over Time
According to our cognitive network model, attention to beliefs should lead to lower
temperature. We expected that temperature of the network would decrease as individuals
took part in the longitudinal survey about their beliefs related to the safety of GM food
and childhood vaccination, because filling in these questionnaires should lead to a higher
amount of attention directed at these beliefs. This higher amount of attention is expected
to result in lower temperature and therefore higher interdependence between beliefs. In
addition, an estimated network temperature that is not infinitely high, or with some
interdependence between beliefs, is necessary for the relationship between energy and
belief change to hold. Systems with temperature lower than infinity are more likely
to move from inconsistent states to consistent states, while networks in infinitely high
temperature have the same probabilities of changing to any state.
    As can be seen in Figure 3a and 3b, the temperatures of both belief networks de-
creased with time, implying that all beliefs became more interdependent during our GM
food and childhood vaccines studies. This confirms our first prediction that attention to
beliefs would lead to a decrease in temperature. The sharpest decrease in the tempera-
ture was observed between the first and second time point, implying that the sharpest
increase in the interdependence between beliefs was observed between the first and sec-
ond measurement. It is noteworthy that simply administering a questionnaire has the
strongest impact on temperature. A relatively low temperature between wave 2a (beliefs
measured before the intervention) and wave 2b (beliefs measured after the intervention)
means beliefs are more likely to move from high energy states to low energy states. In
psychological terms, more attention directed at the beliefs during the intervention re-
sulted in a closer correspondence between dissonance and felt dissonance. This, in turn,
should lead to belief change. Next, we test whether estimated energies indeed relate to
self-reported dissonance.

Energy and Self-reported Dissonance
According to our model, individual belief network energies should be positively related
to self-reported feelings of dissonance. We calculated individual energies by summing
each individual’s actual beliefs weighted by the estimated couplings (See Methods for
more details on their calculations). We measured self-reported felt dissonance in each
of the three waves (but only one time in wave 2, after the educational intervention)
and correlated them with the energy based on the beliefs at the same time point. We
expected a positive relation between energy and self-reported dissonance. Indeed, with

                                            9
(a)                        GM food            (b)                       Childhood vaccines
                0.31

                                                              0.30
                0.28

                                                              0.26
  Temperature

                                                Temperature
                0.25

                                                              0.22
                0.22

                                                              0.18
                       W1   W2a    W2b    W3                         W1      W2a     W2b       W3

                            Time points                                       Time points

Figure 3: Changes in estimated temperature of belief networks through time for GM
food (a) and childhood vaccines (b), N=979.

some attention directed to the subject at hand, individuals should be sensitive to beliefs
that might be contradictory from one another and report feelings of being uncomfortable,
uneasy, and bothered [23, 30].
    The prediction that belief network energies and self-reported dissonance correlate
positively received support among some groups of participants. The relationship between
energies and self-reported dissonance in each wave, separating participants who first held
positive or negative beliefs about GM food or childhood vaccines, is shown in Figure 4a-f.
Regarding beliefs toward GM food, considering all participants together, the correlations
between belief network energies and dissonance did not follow a clear pattern and were
mostly of weak magnitude. However, when considering participants with negative beliefs
and those with positive beliefs separately, we found some interesting trends. There
was a positive relationship between belief network energies and self-reported dissonance
for participants holding somewhat positive beliefs about either topic. The relationship
did not hold for participants with negative beliefs. Regarding beliefs toward childhood
vaccines, considering all participants together, the correlations between belief network
energies and dissonance were positive and of moderate magnitude. However, similar to
the beliefs toward GM food, this relation was only found for individuals with positive
beliefs.
    We believe that the lack of relationship between estimated energies and self-reported
dissonance for participants who generally hold negative views towards GM food or vac-
cines is due to our measurement of felt dissonance, which might have been to unspecific,
and its relationship with overall beliefs. Indeed, felt dissonance was strongly influenced
by one’s original beliefs. Across studies and waves (Figure 4a-f), holding negative views
towards vaccines was associated with higher felt dissonance. Participants holding nega-

                                               10
Wave 1                                                         Wave 2                                                             Wave 3
                                                                                                        GM food
              (a)                                                             (b)                                                             (c)
                  7

                                                                                7

                                                                                                                                                    7
                  6

                                                                                6

                                                                                                                                                    6
Felt dissonance
                  5

                                                                                5

                                                                                                                                                    5
                  4

                                                                                4

                                                                                                                                                    4
                  3

                                                                                3

                                                                                                                                                    3
                  2

                                                                                2

                                                                                                                                                    2
                  1

                                                                                1

                                                                                                                                                    1
                           −0.04     −0.03     −0.02     −0.01     0.00                    −0.04    −0.03     −0.02    −0.01      0.00                       −0.04     −0.03     −0.02     −0.01     0.00

                      r = −0.22 *** (N.B.: r = −0.26 ***, P.B.: r = 0.13 )          r = −0.12 ** (N.B.: r = −0.23 ***, P.B.: r = 0.37 *** )              r = −0.03 (N.B.: r = −0.09 , P.B.: r = 0.33 *** )

                                                                                               Childhood vaccines
              (d)                                                             (e)                                                             (f)
                  7

                                                                                7

                                                                                                                                                    7
                  6

                                                                                6

                                                                                                                                                    6
Felt dissonance
                  5

                                                                                5

                                                                                                                                                    5
                  4

                                                                                4

                                                                                                                                                    4
                  3

                                                                                3

                                                                                                                                                    3
                  2

                                                                                2

                                                                                                                                                    2
                  1

                                                                                1

                                                                                                                                                    1

                           −0.04     −0.03     −0.02     −0.01     0.00                    −0.04    −0.03     −0.02    −0.01      0.00                       −0.04     −0.03     −0.02     −0.01     0.00

                      r = 0.27 *** (N.B.: r = −0.18 *, P.B.: r = 0.36 *** )          r = 0.37 *** (N.B.: r = −0.01 , P.B.: r = 0.38 *** )               r = 0.33 *** (N.B.: r = −0.21 *, P.B.: r = 0.35 *** )

                                                                                                         Energy

Figure 4: Relationship between belief network energies and self-reported felt dissonance,
N=979. Black dots represent individuals who had belief sum scores lower than 0, indicat-
ing negative beliefs. Red triangles represent individuals who had belief sum scores equal
or higher to 0, indicating neutral or positive beliefs. N.B.: Correlation estimates for
individuals holding negative beliefs. P.B.: Correlation estimates for individuals holding
neutral or positive beliefs.

                                                                                                   11
tive beliefs toward GM food and childhood vaccines probably felt dissonance due to their
impression that they were taking part in a study run by individuals with different be-
liefs. Participants were aware that the study was run by scientists, who are likely to hold
positive beliefs toward GM food and childhood vaccines. Therefore, these participants
probably experienced dissonance due to their beliefs being inconsistent with their im-
pression of who created the questionnaire, and not due to the inconsistency of their own
beliefs. While our model was estimated removing overall individual differences in beliefs,
self-reported felt dissonance was inconsistent between people with different beliefs about
GM food and vaccines. This also explains some differences between the GM food and
vaccines study. We only observed a positive correlation between dissonance and energies
for all participants in the childhood vaccines group, probably because there were more
participants holding negative beliefs about GM food than about childhood vaccines.
     To summarise, our prediction that belief network energies relate positively to self-
reported dissonance received mixed support. While this relation was found for individuals
with positive beliefs regarding GM food and childhood vaccines, it did not hold for
individuals with negative beliefs. It is likely that this was caused by the specific phrasing
of the dissonance questions.

Energies Predicting Belief Change
Finally, our model predicts that individual energies should predict belief change and belief
change should lead to lower energies. Most of our participants changed their beliefs over
time, but not always in the expected direction. We measured belief change in moral beliefs
by comparing the mean of all moral beliefs pre- and post-experimental intervention. We
focused on moral beliefs because they reflect participants’ own opinions and attitudes
towards GM food and childhood vaccines to a larger extent than their social beliefs.
In our sample, 46% on average had changes in their networks towards more accepting
(positive) beliefs regarding GM food and childhood vaccines. Even though the education
intervention showed evidence for the safety of GM food and childhood vaccines, 42%
of our participants changed their beliefs on average towards more scepticism (negative
beliefs). This type of backlash is quite common in studies of beliefs about GM food
and childhood vaccines [19, 43]. Beliefs regarding GM food were more likely to change
negatively compared to beliefs regarding childhood vaccines. However, according to our
cognitive network model (Figure 1), the relationship between network energies and belief
change should hold regardless of the direction of belief change.
     In line with our prediction, we find that energies of belief networks estimated before
the interventions can predict which individuals are most likely to change their beliefs
during the interventions. In the study, participants were divided into five experimental
groups for the GM food study and four experimental groups for the childhood vaccines
study (and one control condition in each study, where participants did not receive any
intervention; we excluded this control condition from the current analysis), each of which
received the same scientific message about safety with a different framing (for the full
list of educational interventions, see Supplementary Table 2 in Supplementary Materials).
We calculated correlations between estimated energies and belief change separately for

                                             12
(a)                                                   GM foods                          (b)                                              Childhood vaccines

                                1.0

                                                                                                                            0.8
                                0.8
      Absolute belief change

                                                                                                 Absolute belief change
                                                                                                                            0.6
                                0.6

                                                                                                                            0.4
                                0.4

                                                                                                                            0.2
                                0.2
                                0.0

                                                                                                                            0.0
                                         −0.030    −0.022       −0.014     −0.006     0.002                                       −0.039   −0.029      −0.019   −0.009     0.001
                                                            Energies                                                                                 Energies

                               Information               Farmers              Scientists                                  Tradition              Simple            Big corporations

      (c)                                                                                  GM foods
                                −0.004
      Energies
                                −0.007
                                −0.010

                                              Information                Farmers              Scientists                                   Tradition             Simple

      (d)                                                                            Childhood vaccines
                                −0.009
      Energies
                                −0.012
                                −0.015

                                                  Information                 Scientists                                              Simple               Big corporations

Figure 5: Relationships between belief network energies before interventions and absolute
change of beliefs during interventions, N=979. (a) and (b) show the relation between
energies and absolute belief change for GM foods, and childhood vaccines, respectively,
where each dot represents a participant. (c) and (d) show the belief network energies
before (lighter saturation of bar colours) and after (darker saturation of bar colours)
the interventions. (a) and (b) show that belief network energies correlate with whether
individuals will change their beliefs during interventions. (c) and (d) show that these
changes in beliefs are driven towards lower energy states. Colours of points and bars
correspond to interventions. Error bars in (c) and (d) indicate means ± 1 standard
error.
                                                                                              13
GM food and childhood vaccines for each experimental intervention group in order to
account for potential differences in responses to the framing. Meta-analyses combining all
interventions showed weak positive correlations between energies and absolute change for
both GM food (see Figure 5a), r = .12, p = .01, and childhood vaccines (see Figure 5b),
r = .16, p = .004 (see Supplementary Table 4 for correlations per experimental group).
Individuals with higher belief network energies were more likely to change their beliefs
compared to individuals with low belief network energies. In other words, individuals
who had less consistent beliefs before the intervention were more likely to change their
beliefs after the intervention. In contrast, individuals who had more consistent beliefs
were unlikely to change their beliefs.
    As predicted by our cognitive network model, individuals who changed their beliefs
were more likely to move to lower energy states. We tested whether individuals’ belief
network energies were lower after the intervention compared to before the intervention.
Meta-analyses combining all intervention groups showed that indeed energies were lower
after the interventions for both GM food (see Figure 5c), d = .34, p < .001, and childhood
vaccines (see Figure 5d), d = .33, p < .001 (see Supplementary Table 5 for differences per
experimental group). These effects were weak to moderate. Taken together, these anal-
yses indicate that interventions aiming at changing individuals’ beliefs lead to reconfigu-
ration of beliefs allowing individuals to move to lower energy states and more consistent
belief networks. These re-configurations, however, can be either in line with the inter-
vention’s aim or a backlash. In additional analyses, we found that reduction of energy
is associated with absolute belief change even when controlling for other individual-level
characteristics across educational interventions (Supplementary Materials Figure 5).

Discussion
In this paper, we showed that formalising dissonance as energy of a belief network can
be useful to predict and understand mechanisms leading to belief change. We proposed
a cognitive network model which combines both social and moral beliefs and tested it
using a longitudinal survey. This model enabled us to precisely estimate important pre-
dictors of belief change, such as the relationship between beliefs (couplings), their overall
dissonance (energy), and the randomness of beliefs (temperature). Expanding on the
AE framework [39], we estimated cognitive networks combining social and moral beliefs.
Using a longitudinal nationally representative study, we found full or partial support for
our three predictions derived from the model. First, in line with our prediction, we found
that attention to beliefs during the study was associated with a decrease in estimated
temperature of the network. Second, we found partial support for our prediction that
individual estimated energies should positively relate to self-reported feelings of disso-
nance. This relation was found for participants with positive beliefs towards GM food
or childhood vaccines. Finally, we presented evidence in support of our third prediction:
Individual-level energies were related to belief change after an educational intervention
and this belief change lead to lower energy states. While our cognitive network model
is only an analogy for actual cognitive processes, these findings show its usefulness in

                                             14
estimating and disentangling key psychological factors influencing belief change.
    We have two main contributions. Our first main contribution is to combine social
and moral beliefs into a single cognitive network model built through a statistical physics
framework. This model extends our recent framework for unifying moral and social beliefs
[22] by also taking the network structure of all beliefs into account. Additionally, this
model draws on previous research combining moral and social beliefs [34] and network
models on relationships between beliefs [35, 36]. Previous research on belief formation
and change have stressed the importance of both these sets of factors as individuals make
decisions. Due in part to lack of cross-disciplinary research, however, the combination of
both sets in one framework remains rare. In this paper, we draw on social psychology and
statistical physics to not only incorporate beliefs across these two domains, but include
them as part of an interacting network. We hope this research encourages more studies
of the interactions between social and moral belief networks as important determinants
for belief change.
    Our second main contribution is that our cognitive network model is able to empir-
ically predict belief change by connecting physical parameters to actual psychological
constructs. Many belief dynamic models have remained untested on empirical data. In
addition to a formal model, we provide empirical predictions about belief change using
data collected specifically to answer these questions. Using a model based in social psy-
chology, we bridge the gap between belief dynamics models in statistical physics and
empirical work on science communication. We develop clear psychological meanings for
statistical physics parameters and test their empirical validity. Belief network energies
provide a formalisation of dissonance and temperature provides a formalisation of atten-
tion directed at an issue. This enables us to illuminate some of the mechanisms behind
belief change: Individuals are motivated to reduce dissonance between beliefs and recon-
figure their beliefs to allow lower dissonance. Such reconfiguration can be, but is not
necessarily, in line with the aim of the intervention. The direction in which individuals
change their beliefs does not only depend on the intervention but also on the easiest way
for individuals to reduce their dissonance. This finding also goes above and beyond the
classic finding that inducing dissonance leads to belief change [40–42] by showing that
providing individuals with new information interacts with dissonances in their belief net-
work. Individuals with low dissonance are unlikely to change at all, while individuals
with high dissonance can change in both directions.
    There are some limitations to the study. First, we estimated temperature per time
point for the whole group of participants because current network estimation methods
are not able to estimate temperature separately for each individual. The group-level
network temperature thus likely represents the average temperature of the group with
individual variation possibly captured by variations in energy. A longer longitudinal
study and more advanced methods would enable individual-level estimates for temper-
ature. Second, we did not have an empirical measure of attention and so could only
infer that our estimated measure of temperature was related to attention through other
proxies. However, temperature could reflect many psychological processes leading to the
likelihood of belief networks moving to more consistent states or not. Third, as discussed

                                            15
above, our model predicted absolute belief change but not the direction of belief change,
towards either acceptance or rejection of the safety of GM food and vaccines. Future
research should expand on this model to provide ways to explain why some individuals
accept or reject an experimental intervention and if individuals are in fact choosing the
“easiest” path to a more consistent belief network. Finally, we focused on cognitive be-
liefs of one individual at a time, however, individuals are connected within larger social
networks which influence the dynamics of belief change over a large population. We hope
that subsequent research will continue to bridge social psychology and statistical physics
to model and test belief change at the individual and societal level.
     This research has implications for science communication regarding issues critical to
the health of many. We expect that scientific educational interventions that focus on
reducing the belief network’s dissonance will be more effective in changing the minds
of science sceptics. This applies specifically to the case of beliefs about GM food and
vaccines but can be expanded to many other scientific issues, such as climate change. This
study shows that given enough attention to the issue, individuals do change their mind
if this enables less dissonance between all their beliefs within their cognitive network.
Science communication should take into account how different moral and social beliefs are
connected to each other to draft educational interventions that could lower the dissonance
of the belief network in a way that leads to more acceptance of scientific facts. Further
investigations to translate our findings to science communication might help combating
erroneous and socially-detrimental beliefs.

Methods
Study Design
We conducted a longitudinal study with an experimental component over three waves
on a probabilistic national sample in the United States. To select participants for the
study, we screened N=2,482 participants for their beliefs about the safety of GM food
and childhood vaccines. We selected N=1,832 individuals who were somewhat hesitant
about the safety of GM food or vaccine for the main experimental study. In other words,
we only included individuals who selected a number between 1 and 6 (included) for
the screener question “Do you think it is unsafe or safe to eat GM food?” or “Do you
think childhood vaccines are unsafe or safe for healthy children?” with the options from
1-completely unsafe to 7-completely safe.
    Of the 1,832 selected participants, 979 completed the three waves with no missing
values on any relevant questions and we only included these participants who had no
missing values in our analyses. The first wave, on average 90 days after the screener,
questioned participants about their beliefs about the safety of GM food and childhood
vaccines as well as about related moral concerns and perceived beliefs of social contacts
and sources. These questions were then administered again in wave 2, on average 20
days later, both before and after an experimental manipulation, and again in wave 3, on
average 10 days later.

                                           16
Questionnaire and Intervention
To measure individuals’ moral and social beliefs about GM food and childhood vaccines,
we included questions about related moral beliefs [44], and participants’ perception of the
beliefs of relevant social groups [45]. Haidt [46] identifies six moral foundations relevant
for different groups of U.S. Americans: Care, Fairness, Loyalty, Authority, Purity, and
Liberty. We developed two questions for each of the moral foundations. For the social
network, we focused on perceived beliefs about the safety of GM food or vaccines from
direct social contacts (family and close friends, online community) and relevant sources
of information (medical doctors, scientists, governmental agencies, online influencers,
journalists, and the US general public). Full list of questions focused on moral and social
beliefs are in Supplementary Table 1.
    We included other questions relevant for the model in each wave of the questionnaire.
First, we developed three questions focused on felt dissonance. These questions asked if
the participant felt at ease, unbothered, and comfortable (all also on a scale from 1 to 7
and recoded so that higher values indicate higher dissonance). We averaged these three
items into an index of felt dissonance. Cronbach’s alphas in the different waves were high
for both GM foods and childhood vaccines (GM foods wave 1: .93, wave 2: .93, wave 3:
.94; Childhood vaccines wave 1: .92, wave 2: .93, wave 3: .95), indicating high reliability.
    In the second survey wave, participants were randomised into different experimental
groups that received scientific facts about GM food and vaccines combined with messages
targeting different social and moral considerations. The Supplementary Materials include
the experimental conditions for participants selected for the GM food study (N=549) and
the childhood vaccines study (N=430) (see Supplementary Table 2). Each message within
the GM and vaccines surveys had similar levels of readability and word count.

Network Estimation
We estimated belief networks including moral and social beliefs for GM food and child-
hood vaccines, respectively. Before estimating the networks we regressed each belief on
person and time to partial out these effects. We then used the residuals of these regres-
sion analyses to estimate the networks. We implemented our theoretical model using the
Gaussian Graphical Model (GGM), which is the most common approach to estimate net-
works from continuous data. Edges in a network represent partial correlations between
two nodes while controlling for all other nodes. Modelling the variance-covariance matrix
Σ can be done in the following way [47]:

                                    Σ = ∆(I − Ω)−1 ∆                                    (3)

where Ω represents the partial correlations between nodes and measures the couplings ω
of our cognitive network model. ∆ represents a diagonal scaling matrix with square roots
of the diagonal precision matrix scaling the partial correlations on the diagonal and 0s
on the off-diagonal. These scaling values measure the temperature β1 of our model. The
difference between these scaling values and temperature is that there is one scaling value
for each belief, while there is a single value for temperature in our model. The reason to

                                            17
have a separate scaling value for each belief is that scaling a GGM by a single value often
results in a variance-covariance matrix that is not positive definite. As is the case for
temperature, lower scaling values result in higher correlations between beliefs, because
the model-implied correlations result from dividing the partial correlation between two
given beliefs by the product of their scaling values. The average of these scaling values
can therefore be regarded as a measure of temperature.
     We fitted networks separately for GM foods and childhood vaccines across the differ-
ent time points and increasingly constrained the parameters of the specifications of our
model in several steps and assessed the fit of these different specifications based on the
Bayesian Information Criterion (BIC). These specifications were estimated using the R-
package psychonetrics [48]. We compared the fit of eight specifications of our model with
increasing constraints of the estimated networks. We first let all parameters vary freely
across time points and subsequently constrained the following parameters to be equal
across time points: partial correlations between nodes (Ω), intercepts (mean values) of
the nodes, and scaling values (∆, as a proxy of temperature). We include constraints in
the intercepts, because this allows us to use an approach similar to testing measurement
invariance and makes variations in the scaling values identifiable. We tested each con-
straint using either a dense (all nodes being connected) or sparse network (some couplings
set to 0). We determined which couplings were set to 0 using a prune-step-up procedure,
which sets a given coupling to 0 and tests whether this results in better or worse model
fit. We then selected the best fitting specification of the model.
     For both the GM food and the childhood vaccines networks, the best fitting specifica-
tion of our model was a sparse model (i.e., some partial correlations between beliefs were
set to be 0) with equal partial correlations across time points (i.e., partial correlations
between all beliefs were set to the exact same values at every time point) and intercepts
(mean values) but unconstrained temperature across time points (see Supplementary
Table 3 for fit measures of the different specifications of the model), implying that the
network structure remained constant over time, while temperature varied over time.

Calculation of belief network energies
To calculate belief network energies per person at each time point, we used estimated
partial correlations of the belief network. We multiplied the partial correlation between
any given two beliefs with the recorded responses each individual had on these beliefs.
The belief network energy is then the sum of these pairwise energy scores multiplied by
-1.

Test of Assumptions and Validation
To test the appropriateness of our assumption that Gaussian distributions were fitting
our data, we investigated whether the multivariate distributions of the measured beliefs
confirmed to a normal distribution. As can be seen in Supplementary Figure 1 this
assumption was met. To test the appropriateness of our assumption that couplings es-
timated at the group level were representative of couplings at the individual level, we

                                            18
first investigated the relationship between individual and group variances. The results of
this analysis are shown in Supplementary Figure 2. We found that questions with high
individual-level variance over time tended to also have higher group-variation at one time
point. Second, we compared correlations between beliefs estimated over all time points
without taking into account the multi-level nature of our data with correlations control-
ling for the multi-level nature of our data (with time points nested within individuals).
As can be seen in Supplementary Figure 3 the relative size of correlations between beliefs
was remarkably similar between the different forms of estimations, indicating that group-
level correlation were similar to individual-level correlations. This was also confirmed by
almost perfect correlations between the different estimation methods (GM foods: r (188)
= .98, p < .001; childhood vaccines: .97 (188), p < .001). Only the absolute size of
the correlation coefficients was affected by the different estimation methods. Correlation
coefficients were considerably higher when the multi-level nature of the data was not
taken into account (GM food: mean r = .39; childhood vaccines: mean r = .41) than
when it was taken into account (GM food: mean r = .24; childhood vaccines: mean r =
.28). Taken together, we conclude that it is likely that couplings estimated at the group
level were representative of couplings at the individual level.
     In order to further test the validity of the model, we compared estimated and self-
reported centrality of beliefs. Investigating the relationship between estimated and self-
reported centrality allowed us to test whether the estimated network structure is in line
with the subjective perception of the participants. For this analysis, we made use of
additional questions in our data. Participants rated how important their different beliefs
are for their belief about the safety of GM food or childhood vaccines. Participants also
rated to what extent they believed that GM food or childhood vaccines are safe. For
this analysis, we re-estimated the belief network including the safety beliefs. Results of
this analysis are shown in Supplementary Figure 4. We found that estimated measures
of centrality in relation to the safety of GM food and childhood vaccines correlated with
self-reported importance of moral and social beliefs.

Meta-Analyses on Energies and Belief Change
Correlation between energies and absolute belief change
To test whether belief network energies predict belief change, we first correlated belief
network energies and absolute belief change separately for each intervention and for
each topic (these correlations and their associated significance levels can be found in
Supplementary Table 4). We then transformed these Pearson’s correlation into Fisher’s
z-scores and entered the scores into a random-effects meta-analyses separately for GM
food and childhood vaccines. Finally, we re-transformed the Fisher’s z-scores back to
correlation coefficients for ease of interpretation.

                                            19
Differences between energies before and after the interventions
To test whether belief network energies decrease after the interventions, we first calculated
the mean differences between energies before and after each intervention separately for
each topic (these mean differences and their associated significance levels can be found
in Supplementary Table 5). We then transformed these scores into standardised mean
change scores and entered the scores into a random-effects meta-analyses separately for
GM food and childhood vaccines. Finally, we re-transformed the standardised mean
change scores back to raw differences for ease of interpretation.

References
 [1] World Health Organization, “Ten threats to global health in 2019 Air pollution and
     climate change,” pp. 1–18, 2019.

 [2] C. Funk, B. Kennedy, and M. Hefferon, “Vast majority of Americans say benefits of
     childhood vaccines outweigh risks,” tech. rep., Pew Research Center, 2017.

 [3] World Health Organization, “Malnutrition,” 2020.

 [4] Committee on Genetically Engineerd Crops: Past Experience and Future Prospects;,
     Board on Agriculture and Natural Resources;, and Division on Earth and Life Stud-
     ies;, “Genetically Engineered Crops: Experiences and Prospects,” tech. rep., The
     National Academies of Sciences, Engineering, Medicine, Washington, DC, 2016.

 [5] C. Funk, L. Rainie, and A. Smith, “Public and Scientists’ Views on Science and
     Society,” Pew Research Center, pp. 1–18, 2015.

 [6] C. Funk and B. Kennedy, “The New Food Fights: U.S. Public Divides Over Food
     Science,” tech. rep., Pew Research Center, 2016.

 [7] J. Dalege, D. Borsboom, F. Van Harreveld, H. van den Berg, M. Conner, and H. L. J.
     van der Maas, “Toward a formalized account of attitudes: The Causal Attitude
     Network (CAN) Model,” Psychological Review, vol. 123, no. 1, pp. 2–22, 2016.

 [8] B. M. Monroe and S. J. Read, “A General Connectionist Model of Attitude Structure
     and Change: The ACS (Attitudes as Constraint Satisfaction) Model,” Psychological
     Review, vol. 115, no. 3, pp. 733–759, 2008.

 [9] F. Van Overwalle and F. Siebler, “A Connectionist Model of Attitude Formation
     and Change,” Personality and Social Psychology Review, vol. 9, no. 3, pp. 231–274,
     2005.

[10] A. B. Amin, R. A. Bednarczyk, C. E. Ray, K. J. Melchiori, J. Graham, J. R.
     Huntsinger, and S. B. Omer, “Association of moral values with vaccine hesitancy,”
     Nature Human Behaviour, vol. 1, no. 12, pp. 873–880, 2017.

                                             20
[11] S. Clifford and D. G. Wendell, “How Disgust Influences Health Purity Attitudes,”
     Political Behavior, vol. 38, no. 1, pp. 155–178, 2016.

[12] D. M. Kahan, H. Jenkins-Smith, and D. Braman, “Cultural cognition of scientific
     consensus,” Journal of Risk Research, vol. 14, no. 2, pp. 147–174, 2011.

[13] H. Miton and H. Mercier, “Cognitive Obstacles to Pro-Vaccination Beliefs,” Trends
     in Cognitive Sciences, vol. 19, no. 11, pp. 633–636, 2015.

[14] B. T. Rutjens, R. M. Sutton, and R. van der Lee, “Not All Skepticism Is Equal:
     Exploring the Ideological Antecedents of Science Acceptance and Rejection,” Per-
     sonality and Social Psychology Bulletin, vol. 44, no. 3, pp. 384–405, 2018.

[15] S. E. Scott, Y. Inbar, and P. Rozin, “Evidence for Absolute Moral Opposition to Ge-
     netically Modified Food in the United States,” Perspectives on Psychological Science,
     vol. 11, no. 3, pp. 315–324, 2016.

[16] N. F. Johnson, N. Velásquez, N. J. Restrepo, R. Leahy, N. Gabriel, S. E. Oud,
     M. Zheng, P. Manrique, and S. Wuchty, “The online competition between pro- and
     anti-vaccination views,” Nature, no. January, 2020.

[17] C. Betsch, C. Ulshöfer, F. Renkewitz, and T. Betsch, “The influence of narrative
     v. statistical information on perceiving vaccination risks,” Medical Decision Making,
     vol. 31, no. 5, pp. 742–753, 2011.

[18] Z. Horne, D. Powell, J. E. Hummel, and K. J. Holyoak, “Countering antivaccina-
     tion attitudes,” Proceedings of the National Academy of Sciences, vol. 112, no. 33,
     pp. 10321–10324, 2015.

[19] B. Nyhan, J. Reifler, S. Richey, and G. L. Freed, “Effective Messages in Vaccine
     Promotion: A randomized Trial,” Pediatrics, pp. 11–18, 2014.

[20] A. J. Rothman and P. Salovey, “Shaping perceptions to motivate healthy behavior:
     The role of message framing,” Psychological Bulletin, vol. 121, no. 1, pp. 3–19, 1997.

[21] E. Dubé, D. Gagnon, N. E. MacDonald, J. Eskola, X. Liang, M. Chaudhuri, E. Dube,
     B. Gellin, S. Goldstein, H. Larson, M. L. Manzo, A. Reingold, K. Tshering, Y. Zhou,
     P. Duclos, S. Guirguis, B. Hickler, and M. Schuster, “Strategies intended to address
     vaccine hesitancy: Review of published reviews,” Vaccine, vol. 33, no. 34, pp. 4191–
     4203, 2015.

[22] M. Galesic, H. Olsson, J. Dalege, T. van der Does, and D. L. Stein, “Integrating
     Social and Cognitive Aspects of Belief Dynamics: Towards a Unifying Framework.”
     2020.

[23] A. J. Elliot and P. G. Devine, “On the Motivational Nature of Cognitive Disso-
     nance: Dissonance as Psychological Discomfort,” Journal of Personality and Social
     Psychology, vol. 67, no. 3, pp. 382–394, 1994.

                                            21
You can also read